From Data-Centric to Sample-Centric: Enhancing LLM Reasoning via Progressive Optimization

TL;DR

This paper introduces LPPO, a progressive optimization framework for large language models that leverages high-quality demonstrations and dynamic sample weighting to improve reasoning capabilities efficiently.

Contribution

The paper proposes LPPO, a novel sample-centric approach combining prefix-guided sampling and learning-progress weighting to enhance LLM reasoning with limited high-quality data.

Findings

Outperforms strong baselines on mathematical reasoning benchmarks.

Achieves faster convergence and higher performance ceiling.

Effectively leverages expert demonstrations for improved reasoning.

Abstract

Reinforcement learning with verifiable rewards (RLVR) has recently advanced the reasoning capabilities of large language models (LLMs). While prior work has emphasized algorithmic design, data curation, and reward shaping, we investigate RLVR from a sample-centric perspective and introduce LPPO (Learning-Progress and Prefix-guided Optimization), a framework of progressive optimization techniques. Our work addresses a critical question: how to best leverage a small set of trusted, high-quality demonstrations, rather than simply scaling up data volume. First, motivated by how hints aid human problem-solving, we propose prefix-guided sampling, an online data augmentation method that incorporates partial solution prefixes from expert demonstrations to guide the policy, particularly for challenging instances. Second, inspired by how humans focus on important questions aligned with their current capabilities, we introduce learning-progress weighting, a dynamic strategy that adjusts each training sample's influence based on model progression. We estimate sample-level learning progress via an exponential moving average of per-sample pass rates, promoting samples that foster learning and de-emphasizing stagnant ones. Experiments on mathematical-reasoning benchmarks demonstrate that our methods outperform strong baselines, yielding faster convergence and a higher performance ceiling.